Eidetic memory profile

Type: agentic-system-analyses/types/agentic-system-memory-profile.md

Comparison rationale

This profile classifies the accepted records at the supplied frozen revision. All positive implementation findings have a wired basis, except the explicitly claimed user-only base policy and afforded human editing/browsing surfaces. There is no observed-run or causal classification. Every axis remains partial: its inventory retains recorded initialization, opaque alternative-consumer or provider limits rather than turning a local witness into a complete system list. Provider inference components identify the unresolved boundary; they do not establish that pretrained weights are accumulated memory.

Reconciliation supersedes EPI-CMP-vault-models with MEM-CMP-export-models, and EPI-OBJ-confidence-events with MEM-OBJ-evidence-events. Citations use those canonical declarations. Card containers remain distinct from M3 provenance, M2 managed regions, indexes, generated rules and vault pages; no other supersession changes a value.

Dense retrieval parts map to parametric form on a SQLite substrate; localized confidence numbers map to symbolic form. The known embedding derivation is compiled lineage, not a finding about unknown initial card provenance or external weight history. See MEM-OBJ-memory-chunks and MEM-OBJ-evidence-events.

Admission is classified separately from authoring and I/O. Hook/model gates and export/maintenance software decide generated content automatically. Generic direct caller identity leaves control unresolved, while explicit-user base instructions remain only a claimed manual policy. Human editing is an afforded manual route, not a universal approval barrier. See RT-RTE-explicit-admission, MEM-RTE-card-withdrawal and MEM-RTE-evidence-projection.

Consolidation selects existing claims. Default M2 also deduplicates and revises managed regions; its name does not establish creation of new claims. Optional callback, polish and topic output semantics remain unresolved, so no synthesize value is asserted. A fact-preservation defect would not itself establish synthesis. Promotion/decay/invalidation concern salience or current reliance, not proof or universal deletion. See EPI-RTE-m2-reshaping, EPI-RTE-vault-generation and MEM-RTE-card-withdrawal.

Foreground requested recall is pull; lifecycle delivery is push. Requested returns are not double-counted as push. Automatic maintenance, cache and topic input selection are separately named consumers. Project/session identity matches are targeted selectors; availability/count/budget filters remain coarse. Similarity and relevance evaluators keep their own signal findings. Detail IDs on requested returns alone establish no push signal. See RT-RTE-memory-recall, RT-RTE-context-injection, MEM-RTE-session-hints and MEM-RTE-vault-export.

Trace-learning yes describes automatic trace-fed durable behavior-shaping writes, not improved capacity. Session-derived knowledge receives learning force at its update consumer and knowledge force at its later reader. Cached definitive outcomes also shape later candidate routing; raw diagnostic retention alone does not qualify. Session counts select later guidance. Test-correlated Evidence can shape later ranking, but the ordinary hook project-identity mismatch prevents claiming that ordinary filing receives these promotions. See RT-RTE-signal-compound, RT-RTE-m3-filing, MEM-OBJ-admission-traces, MEM-RTE-session-hints and MEM-RTE-test-affirmation.

Trace sources follow original content. Conversation mining is session-logs; session occurrence history is event-streams; Bash test execution feedback is tool-traces. Adapter event labels do not transform conversations or static source files into event-stream sources. A linked multi-step workflow is not itself evidence that the update consumes trajectories. Missing provenance stays local to unresolved routes, not a system-wide negative.

Source-read log: none. Classification used only the supplied accepted members, types, collection and worker/record contracts. No frozen source ambiguity required a direct source read, no evidence or record was added, and no peer report was changed. Deterministic output validation is the form/reference check performed for this attempt; it does not establish analytical acceptance or activation.